LangExtract

LangExtract is an open-source Python library published by Google under the Apache 2.0 license. It uses large language models to extract structured information from unstructured text documents based on user-defined instructions and few-shot examples, mapping every extraction back to its exact location in the source text. The library is authored by Akshay Goel (goelak@google.com) and is available on PyPI as langextract.

What It Is

LangExtract sits in the structured-data-extraction category of NLP tooling. Its core job is to take free-form text — clinical notes, legal documents, literary works, radiology reports — and produce typed, schema-consistent output objects that are traceable to specific character intervals in the original document. Unlike generic LLM prompting, LangExtract enforces output schemas, handles long-document chunking automatically, and generates self-contained interactive HTML files for reviewing thousands of extracted entities in context. It requires no model fine-tuning; a few high-quality examples are enough to define a new extraction task.

Core Capabilities

• Precise source grounding: Every extraction is mapped to a char_interval in the source text. Extractions that cannot be located in the source (e.g., hallucinated from few-shot examples) are flagged with char_interval = None and can be filtered out.
• Reliable structured outputs: Output schemas are enforced using controlled generation on supported models (Gemini) or structured outputs / JSON mode on OpenAI models.
• Long-document optimization: Text chunking, parallel processing (max_workers), and multiple extraction passes (extraction_passes) address the "needle-in-a-haystack" problem for documents of arbitrary length.
• Interactive visualization: A single lx.visualize() call generates an animated, self-contained HTML file for reviewing entities in their original context.
• Flexible LLM support: Built-in providers for Google Gemini (via AI Studio or Vertex AI), OpenAI (GPT-4o, GPT-4o-mini, and compatible endpoints), and local models via Ollama. A plugin system allows community-contributed providers.
• Domain agnostic: Demonstrated use cases include literary character extraction, medication and dosage extraction from clinical notes, and radiology report structuring (RadExtract demo on HuggingFace Spaces).

Model Support and Provider Architecture

LangExtract ships with three built-in providers and a lightweight plugin system for extending to additional backends:

• Gemini (default): gemini-3.5-flash is the recommended default. Vertex AI Batch API is supported for cost-optimized large-scale workloads.
• OpenAI: Requires pip install langextract[openai]. Supports structured outputs, JSON mode, and the OpenAI Batch API for high-volume jobs. Compatible with any OpenAI-compatible endpoint via ModelConfig.
• Ollama (local inference): Runs models like gemma2:2b locally without an API key. Uses JSON mode via FormatModeSchema.
• Custom providers: Registered via @router.register(...) from langextract.providers, distributable as separate Python packages, and resolved via priority-based routing.

Setup Path

Installation is a single pip command (pip install langextract) requiring Python ≥ 3.10. Optional extras include [openai], [dev], [test], and [notebook]. Docker support is included for containerized workflows. API keys for cloud models can be supplied via environment variable (LANGEXTRACT_API_KEY), a .env file, direct parameter, or Vertex AI service accounts.

Update: v1.6.0 (July 2, 2026)

The current release is v1.6.0, published July 2, 2026. The release history shows rapid iteration since the initial 1.0.0 release in July 2025, with major versions reaching 1.6 within roughly one year. Notable milestones in the release history include v1.1.0 (November 2025), v1.2.0 (March 2026), v1.3.0 (April 2026), v1.4.0 and v1.5.0 (May 2026), and v1.6.0 (July 2026). The project carries a Zenodo DOI (10.5281/zenodo.17015089), indicating it is intended for academic citation. The GitHub repository is tagged with topics including gemini, gemini-api, information-extraction, llm, nlp, and structured-data.